Conversation Analytics: understand what your visitors ask

The analytics view delivers 32 metric cards in one response, plus a daily series, topics, unanswered questions, knowledge gaps, blocked cases, drop-off points, recommendations and a version comparison. Things are counted, not estimated: a card without a basis in data carries the note not measured together with a reason instead of a 0, and every share shows its numerator and denominator beside it. Recorded are, among others, conversations, messages, self-resolved cases, handovers, answers with a citation, enquiries, appointments, carts and drop-offs. The distribution knows two attributes, language and window class — no browser identification is read for it. A revenue figure or an attribution to a purchase is deliberately absent. More under conversation analytics.

In most cases the source is missing, not the capability. An answer is built from the sections that were found: at most five search hits plus your text blocks go to the model, numbered, and sourced answers name the excerpts they rely on. If the assistant finds nothing suitable, it usually says that nothing on this is held in the knowledge base. That refusal, like an answer without a used excerpt, is recorded as a gap; from the Pro package you will find it again in conversation analytics among the detected gaps. The route is therefore clear: read the gap, put the content into the knowledge base, re-read the source.

Both, but kept apart. Analytics shows condensed metrics and topics so you can recognise patterns without profiling individual visitors; it evaluates your own conversations only and knows neither an industry benchmark nor revenue attribution. You can look at individual conversations — masked by default: email address, phone, IBAN and postal address appear as readable placeholders, and the response itself states that it is masked. The plain text is seen only by someone who holds the role for it, states a reason and accepts the log line. Masking protects the display, not the storage: the history remains stored in full.

It shows which products are asked about in the chat and where conversations break off. Topics are formed from the most recent 5000 visitor questions in the selected period, and the list ends at 15 entries. A drop-off is derived rather than guessed: an open conversation, at least one visitor message, silent for thirty minutes, not handed over and without an enquiry, appointment or cart. On top of that come recommendations in ten kinds, each with counted cases and the area in which you make the change. What the evaluation does not do: it calculates no revenue value and attributes no amount. How the shop assistant carries out the resulting actions is described on its page.

Conversation analytics belongs to the Pro package: one-off setup from 2,900 € net, operation from 149 € per month net, including shop connection, product cards and the cart in the chat. The analytics endpoints sit behind that plan boundary, not behind an extra line item. Inside you see metric cards, daily curves, topics, unanswered questions, knowledge gaps, abandonment points and a comparison of two versions of your assistant, across periods of 7, 30 and 90 days. Every figure is counted, not estimated: where the basis is missing, the field says unmeasured together with a reason, and not a zero. All services can be cancelled monthly, there is no minimum term. Overview under pricing.

The evaluation refreshes itself every 60 seconds and marks a reading as stale from 90 seconds on; the figures behind it sit in the cache for up to 60 seconds. On top of that comes a weekly report compiled early on Monday over the preceding seven days, in the Europe/Berlin zone, with twelve labelled cards ranging from conversations through self-resolved cases to the drop-off share. A card without a basis carries its reason in plain words instead of a zero, so that a failed measurement does not look like a result. The resulting improvements to assistant, knowledge base and actions we implement together with you and then measure the effect.

Yes. Conversation analytics works with aggregated metrics and topics instead of personal profiles. The analysis runs on our servers in Germany, governed by a data processing agreement with clear deletion periods; the answers in the chat itself are generated by a specialised model provider, also outside the EU on the basis of standard contractual clauses. Your conversation data belongs to you and is neither sold nor passed on to third parties. Technically, analytics reads no browser identifier and stores no visitor address, only language and viewport class; the similarity of two topics is computed in PHP on our own server, without an external service and without a model call. Every figure is counted: where the basis is missing, the field says unmeasured together with a reason instead of a zero. Details under privacy and hosting.